基于改进U-Net的低剂量CT图像重建方法  

Method of low⁃dose CT image reconstruction based on improved U⁃Net

作　　者：朱榕榕 王明泉[1] 曹鹏娟 范涛 ZHU Rongrong;WANG Mingquan;CAO Pengjuan;FAN Tao(MOE Key Laboratory of Instrumentation Science and Dynamic Measurement,North University of China,Taiyuan 030051,China)

机构地区：[1]中北大学仪器科学与动态测试教育部重点实验室,山西太原030051

出　　处：《现代电子技术》2023年第9期41-45,共5页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(61171177);国家重大科学仪器设备开发专项(2013YQ240803)。

摘　　要：针对低剂量CT图像重建会产生噪声和伪影的问题,在U-Net神经网络基础上引入残差学习和空间注意力机制,在编解码过程中嵌入跳跃连接为上采样增加多尺度信息,使用AAPM公开数据集CT影像进行模型训练和测试。选取峰值信噪比(PSNR)、结构相似性(SSIM)和均方根误差(RMSE)作为图像性能评价指标。在CT重建结果的测试中,与未处理的图像相比,网络模型处理后图像的PSNR、SSIM和RMSE指标平均值分别提升21.699%、2.263%和40.833%。实验结果表明,改进的U-Net神经网络模型能够减少噪声和伪影,保留了更多的纹理细节,对低剂量CT重建图像质量的提高有一定效果。To address the problem of noise and artifacts in low⁃dose CT image reconstruction,the residual learning and spatial attention mechanism are introduced based on the U⁃Net neural network.The jump connection is embedded in the process of encoding and decoding to increase the multi⁃scale information for the upsampling,and the model is trained and tested by means of the public AAPM dataset of CT images.The peak signal to noise ratio(PSNR),structural similarity(SSIM),and root mean square error(RMSE)are selected as image performance evaluation indicators.In CT reconstruction testing,in comparison with unprocessed images,the PSNR,SSIM and RMSE indicators of the images processed by the network model is increased by 21.699%,2.263%and 40.833%,respectively.The experimental results show that the improved U⁃Net neural network model can reduce noise and artifacts,preserve more texture details,and has a certain effect on improving the quality of low⁃dose CT reconstructed images.

关 键 词：低剂量CT 图像重建 神经网络 残差网络 空间注意力机制 图像去噪 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP391[电子电信—信息与通信工程]